Europe’s new transparency rules for AI-generated content are already encountering a decidedly ordinary enforcement problem: estate agents can alter property photographs considerably faster than a marketplace can inspect them. Dutch property platform Funda requires substantively AI-altered images to be identified and remain realistic, yet manipulated interiors have continued to appear without the disclosure expected by the platform and Europe’s newly applicable transparency regime.

The examples are less spectacular than the political deepfakes that have dominated debate around synthetic media, but they bring the same underlying question directly into commerce. Photographs can make rooms appear larger, introduce digitally generated furniture, present completed renovations, or alter physical characteristics, changing what a prospective buyer believes they are looking at before deciding whether to view or value a property.

Funda allows estate agents to use AI to demonstrate credible possibilities, such as furnishing an empty room or illustrating how an outdated interior might look after renovation. Its rules become stricter when manipulation changes the underlying reality, prohibiting edits that arbitrarily move doors or windows, conceal defects, distort dimensions, or depict construction that would not realistically be possible.

Where a photograph has been materially altered, Funda requires the change to be made clear in the image. The platform does not pre-screen every photograph, however, and instead investigates reports, asks agents to correct non-compliant material, and can remove a listing where the required changes are not made.

Transparency rules meet marketplace operations

That moderation model now operates alongside Article 50 of the EU AI Act, whose transparency provisions began applying on 2 August. Among other measures, the regime requires disclosure of AI-generated or manipulated image, audio, or video content in circumstances covered by its deepfake provisions, while providers of generative systems face separate requirements around machine-readable marking.

The difference between provider and deployer becomes consequential in a commercial marketplace. A business supplying an image-generation model has obligations around the technology it provides, while an estate agency using AI during its own professional activity can carry responsibilities for how qualifying synthetic material is presented to the person seeing it.

Visible disclosure and embedded technical marking also perform different jobs. Metadata can help platforms or automated systems identify provenance, although it may be lost when an image is resized, compressed, screenshotted, or passed through other software. Where an individual needs to be informed that apparently authentic material has been generated or materially manipulated, hidden provenance data cannot perform that communication by itself.

Property advertising provides a useful test because digital alteration existed long before generative AI. Estate agents have routinely adjusted brightness, perspective, and colour, while virtual staging tools have been used to furnish empty rooms for years. The harder regulatory question is therefore not whether software touched a photograph but whether an apparently authentic image materially changes the impression of the asset being offered.

Detection becomes an operational burden

Funda’s own policy addresses that distinction through the physical reality of the property. Cosmetic visualisation can be permitted when appropriately identified, whereas hiding a defect or making a room appear larger can mislead a buyer about something that cannot be resolved simply by replacing furniture or changing decoration.

Yet publishing rules do not automatically create compliance across thousands of listings supplied by independent businesses. A marketplace must decide how much screening to perform before publication, whether automated detectors are sufficiently reliable to flag suspicious images, how complaints should be investigated, and who carries the cost of reviewing borderline cases.

The problem becomes harder because synthetic-content detectors remain imperfect. A model may fail to recognise an edited photograph or incorrectly classify genuine photography, while technical provenance can disappear as files move through ordinary commercial workflows. Where the professional user fails to disclose an alteration, the marketplace may be left trying to infer whether an apparently plausible photograph originated from a camera, conventional editing software, or a generative system.

The same difficulty extends beyond residential property. Recruitment platforms, retailers, travel marketplaces, insurers, and professional-services businesses all distribute large amounts of imagery supplied by customers or commercial partners, and Europe’s transparency regime has to operate across workflows in which one piece of content can pass through several organisations before reaching the person expected to see a disclosure.

Funda already has rules that recognise many of the risks, so the immediate issue is less the absence of policy than the cost and reliability of enforcing it. Synthetic content can be generated almost instantly, whereas deciding whether an image misrepresents a property, identifying who produced it, and securing a correction remains a much slower process involving technology, professional judgement, and ultimately the behaviour of the estate agent responsible for the listing.

The early Dutch experience suggests that AI transparency will be judged less by the wording of disclosure requirements than by whether they survive ordinary commercial volume. Europe’s rules now have to operate in markets where synthetic content is cheap to produce, difficult to detect consistently, and often only subtly different from the conventional digital editing businesses were already using.